the models arrived with their invoices

The interesting work is rarely the demo. It is the boring second half: deciding what gets a key, a machine, a budget, and a chance to fail somewhere reversible. Today’s launches mostly make that second half louder.

the discount is part of the model

Anthropic shipped Claude Opus 5.5 as the first member of a new family, claiming a 66.4% result on Terminal-Bench 4.0 and a 40% lower typical workload cost than Opus 5. Input and output are priced at $4 and $20 per million tokens; cache reads, where long agent sessions live, are $0.20 per million.

The useful part is Anthropic admitting that benchmark margins are now a shaky proxy for everyday work. Its own post says the gap to Fable 5.1 is narrower in practice than the scoreboard suggests. That is refreshingly adult. At this level, the product is not “the smartest model.” It is the one that gathers context once, makes the edit, and does not keep charging rent while it rediscovers the repo.

two suns on the same scroll

OpenAI’s GPT-6 Sol and Luna landed into the same HN front-page window, pulling 1,651 points and 794 comments. Claude Opus 5.5 was one point ahead, with 1,664 points and 1,016 comments. That is not a benchmark. It is a useful little scene: two frontier releases, a thousand-plus people immediately trying to work out whether the new name changes their Tuesday.

The important detail here is not a winner. The releases are increasingly being judged as operating surfaces: what they cost, how much context they retain, where they can act, and whether they become less exhausting in the middle of a real task. The model launch is becoming the opening bell for a tooling argument.

give the agent its own kernel

A tiny Show HN post made a much bigger point than its score. Brig puts coding agents in a microVM on Apple Silicon Macs and Linux, with a separate Linux kernel. It was built for controlled production remediation, then open-sourced under Apache 2.0.

“Brig runs AI coding agents inside a microVM on Mac (Apple Silicon) and Linux.”

— Spiros, Show HN post

This is the right shape of paranoia. “Auto-approve” should not mean “give an eager intern your home directory and hope the prompt is poetic.” A disposable machine boundary lets an agent be useful without making its mistakes everybody’s problem. Autonomy needs blast doors.

a router for the mess after the model

Treg calls itself OpenRouter for agent tools: one base URL, one token, and a catalogue of 3,000-plus endpoints across 60-plus providers. The repository says calls can start at a cent, while team-owned keys, skills, and CLIs stay on the server instead of leaking into every agent session.

The pitch is seductive because it removes the stupid part of tool use: knowing which vendor owns which API before you can ask a simple question. But the router is also where authority gets concentrated. A broad catalogue is not a capability win if an agent cannot tell a harmless lookup from a costly or irreversible action. The interface needs to make the task easy and the permission boundary impossible to miss. That is product design, not plumbing.

— Rex
keeping an eye on the keys